Abstract:

This dataset contains a subset of LADC passive acoustic system EARS buoys data which was collected in 2015 (data inventory can be found in R4.x261.233:0005) and is used to identify three different species of beaked whales in the Gulf of Mexico. The species of beaked whales examined in this dataset are Cuvier’s beaked whale, Gervais’ beaked whale, and an unidentified species that we labeled "BWG", which stands for Beaked Whale of the Gulf. Recordings were processed using a click detection algorithm. Then unsupervised as well as supervised classification algorithms were evaluated for distinguishing species by echolocation features.

Suggested Citation:

Juliette Ioup, Jack LeBien. 2018. Detection and classification of beaked whale echolocation clicks recorded on bottom-moored EARS buoys in the northern Gulf of Mexico from July-October 2015. Distributed by: GRIIDC, Harte Research Institute, Texas A&M University–Corpus Christi. doi:10.7266/N7QF8R9K

Purpose:

Beaked whale detections and classifications were documented to compare between detection and classification algorithm precision and accuracies.

Data Parameters and Units:

EARS_source_data sheet: amplitude (volts), time (seconds), click ID (sequential number of click detections), source filename (filename format is as follows: 5dddFFFF.BBD where 5 = year 2015, ddd = three-digit julian day, FFFF = hexadecimal index number, BB = buoy number, 01 to 10, D = disk number (0 to 3), species (BWG = beaked whale of the Gulf). table1data sheet: click features (mean, standard error of the mean (SEM), minimum and maximum values reported for Cuvier's, Gervais', and BWG [unidentified beaked whale] whales): 10dB bandwidth (Hz), 20 dB center frequency (Hz), spectral centroid (Hz), 10 dB duration (microseconds), 95% energy duration (microseconds), 95% amplitude duration (microseconds), 95% Teager-Kaiser (TK) amplitude duration (microseconds), higuchi fractal dimension, castiglioni fractal dimension, adapted box fractal dimension, Katz fractal dimension, Shannon entropy (bits), Renyi entropy (alpha =3, bits); script (source code used to generate values), SEM (standard error of the mean). fig2data sheet: amplitude (volts; raw signal), envelope (volts; calculated envelope of the raw signal), time (seconds; click duration estimate based on envelope calculations), click ID (sequential click detection number), EARS source file (see above). fig3data sheet: time (seconds; time stamp from the original raw file), click 1 amplitude (volts; raw signal), click 2 amplitude (volts; generated signal created combining the raw signal with random noise), click 3 amplitude (volts; generated signal created by combining the raw signal with random noise), EARS source file (see above for format), click ID (sequential click detection number). fig4data sheet: values represent the average fraction of correct predictions for each trial with increasing levels of noise. Peak frequency (Hz), 10dB bandwidth (Hz), 20 dB center frequency (Hz), spectral centroid (Hz), 10 dB duration (microseconds), 95% energy duration (microseconds), 95% amplitude duration (microseconds), 95% Teager-Kaiser duration (microseconds), higuchi fractal dimension, castiglioni fractal dimension, adapted box fractal dimension, Katz fractal dimension, Shannon entropy (bits), Renyi entropy (alpha =3, bits), mean SNR (dB; average signal to noise ratio for each trial) trial (sequential trial number with varying levels of noise). fig5data sheet: 10 dB duration (microseconds), 95% energy duration (microseconds). These data for three separate Cuvier's clicks within the same click train are segments of the respective click ID data in EARS_source_data. fig6data sheet: purity (0 to 1; represents the purity of clustering using each of the methods [SOM - self-organizing map, k-means, and spectral clustering]), entropy (represents the entropy of clustering using each of the methods [SOM -self-organizing map, k-means, and spectral clustering]), mean SNR (dB; average signal to noise ratio for each click feature set). These are computed features for each click in EARS_source_data. fig7data sheet: 10 dB duration (microseconds), higucki fractal dimension, spectral centroid (Hz), click ID (sequential click detection number). These are computed features for each click in EARS_source_data. fig8data sheet: mean SNR (dB; average signal to noise ratio for each signal used for classification), Features (training feature sets used for feedforward neural network classification; HFD -Higucki fractal dimension, WPD energy map - wavelet packet decomposition energy map), trial (classification accuracy using feedforward neural networks with each training set for signals of various noise levels), mean (average percent accuracy of classification), stand. dev. (standard deviation of classification accuracy). SNR_data sheet: mean SNR (dB; average signal to noise ratio for raw signal [column B] and generated signals with increasing levels of noise [columns C-F]), SNR (signal to noise ratio for each click in raw signal [column B] and generated signals with increasing levels of noise [columns C-F]), click ID (sequential click detection number)

Methods:

k-means, self-organizing maps and spectral clustering analysis for click categorization and identification

Individual Files: